Web Workers / Servers - CPU Frequency vs Core Count

[BotDamian]

Hello everyone,

I haven't done any testing, because I can't think of a "good test".

 

I do web scraping, run various servers etc and I tried to find something about CPU frequency vs Core Count and there is nothing.

For example 7950X vs 5995WX vs Epyc 9554P / Epyc 7713

GHz is basically how fast the CPU executes tasks, so a 5GHz CPU could for example do the same task 2x faster than a Epyc 7713?

[manikyath]

in theory (not counting for overhead of running multiple threads, etc.) the total performance of a chip is the performance per core, times the number of cores.

i should also note here that there's more than just clock speed to how fast a core is. the amount of cache is important, the memory bandwidth makes a difference, etc.

 

the question you need to ask yourself is, how scalable is your process?

 

if you can essentially endlessly divide the process up into more and more processes doing part of the workload, just look at theoretical multicore benchmarks (passmark is probably a good place to start) and scale your workload according to the number of cores you have.

 

if there is severe diminishing returns (or if your task simply cant scale past N cores) the solution is more nuanced, and you're probably best with around N+2 cores (so you can prioritize the cores you're running on to do exclusively the allocated task, leaving the others for OS stuff).

 

there's also the matter of ROI times, more expensive CPU means it'll take longer (more work processed) to pay for itself.

 

also - perhaps you dont necessarily *want* a faster clock speed CPU, because of potential limits on scraping a particular source, in order to not impact said source. (it's a huge no-no to put significant load on a service for the sake of scraping data)

[BotDamian]

On 7/3/2023 at 10:40 PM, manikyath said:

in theory (not counting for overhead of running multiple threads, etc.) the total performance of a chip is the performance per core, times the number of cores.

i should also note here that there's more than just clock speed to how fast a core is. the amount of cache is important, the memory bandwidth makes a difference, etc.

 

the question you need to ask yourself is, how scalable is your process?

 

if you can essentially endlessly divide the process up into more and more processes doing part of the workload, just look at theoretical multicore benchmarks (passmark is probably a good place to start) and scale your workload according to the number of cores you have.

 

if there is severe diminishing returns (or if your task simply cant scale past N cores) the solution is more nuanced, and you're probably best with around N+2 cores (so you can prioritize the cores you're running on to do exclusively the allocated task, leaving the others for OS stuff).

 

there's also the matter of ROI times, more expensive CPU means it'll take longer (more work processed) to pay for itself.

 

also - perhaps you dont necessarily *want* a faster clock speed CPU, because of potential limits on scraping a particular source, in order to not impact said source. (it's a huge no-no to put significant load on a service for the sake of scraping data)

But I could be using 1 core with 2-3 instances if the said core clocked at 4.5Ghz instead of a server Epyc that does 3Ghz tops.